[Updated on 2025-06-15]
Most ML education teaches libraries and frameworks. This is backwards. The field moves fast enough that knowing “why” is the only durable skill. Here’s what actually matters.
The paradigm that defines modern AI is representation learning at scale. Before deep learning, you engineered features by hand. Now the model learns them from data. This changes everything: more data and compute produce better representations, and at sufficient scale, capabilities emerge that don’t exist in small models [1]. The Chinchilla scaling law reframed the question from “how large” to “how much data per parameter” — the optimal ratio is roughly 20 tokens per parameter, meaning most large models are undertrained by a factor of four or more [2].
The transformer is the architecture that made this feasible [3]. Its key innovation is attention: every token communicates with every other token in parallel. This parallelism is why training can scale — RNNs can’t parallelize across time steps, but attention parallelizes across positions. The decoder-only variant is simple: predict the next token, again and again. But the mechanics — QKV projections, the residual stream, layer normalization — transfer across every model you’ll encounter. Debugging gradient flow through a transformer is the closest thing to a universal skill in this field [4]. For NLP specifically, Jurafsky and Martin’s textbook covers the full pipeline from n-grams to RLHF [5].
Pre-training is increasingly commoditized. The real differentiation comes in post-training: instruction tuning, RLHF, DPO, GRPO [6]. The InstructGPT paper showed you can align a pretrained model to follow instructions without changing its architecture [7]. The DPO paper showed you can do it without reinforcement learning at all — direct preference optimization on pairs of outputs [8]. These techniques define the gap between a raw base model and a useful one. The Tuning Playbook from Google Research codifies the experimental methodology that separates systematic work from lucky runs [9]. For the theory behind these methods, Sutton and Barto’s textbook and David Silver’s lectures remain the definitive resources for RL [10]. Lilian Weng’s blog surveys the research landscape better than most published survey articles [11].
The most intellectually urgent open problem is mechanistic interpretability. We’re deploying models we trained but don’t understand internally. Anthropic’s Transformer Circuits work has shown that language models learn interpretable features, compose them into circuits for specific behaviors, and store knowledge in superposition [12]. Sparse autoencoders can extract these features at scale [13]. Neel Nanda’s writing is the best practical starting point for getting into this subfield [14]. This work determines whether we can trust these models with real responsibility. For the theoretical foundations, Murphy covers probabilistic reasoning and Shalev-Shwartz covers why models generalize [15][16].
Building systems with these models requires different thinking than training them [17]. Papers show what’s possible in ideal conditions. Production shows what’s reliable under real constraints. Retrieval-augmented generation grounds models in external knowledge. Evals measure task-specific behavior, not benchmark scores. The most effective agent architectures use simple, composable patterns over complex frameworks [18]. Lilian Weng’s survey of autonomous agents maps the design space — planning, memory, tool use [19]. The gap between a demo and a deployment is large, and closing it requires different skills than training a better model.
If you want to build before you understand, fast.ai gets you running models in a weekend [20]. Then learn linear algebra and probability — Goodfellow’s deep learning textbook is the standard reference [21]. Then implement a transformer from scratch; building one end-to-end reveals more than reading a hundred papers [22]. Then study alignment [7][8]. Everything else — new architectures, frameworks, model releases — is details on top of this foundation. The field rewards depth in fundamentals over breadth in releases.
References
[1] Kaplan et al., “Scaling Laws for Neural Language Models,” 2020. https://arxiv.org/abs/2001.08361
[2] Hoffmann et al., “Training Compute-Optimal Large Language Models,” 2022. https://arxiv.org/abs/2203.15556
[3] Vaswani et al., “Attention Is All You Need,” 2017. https://arxiv.org/abs/1706.03762
[4] CS231n, http://cs231n.github.io/; 3Blue1Brown, https://www.youtube.com/c/3blue1brown; Jay Alammar, https://jalammar.github.io/
[5] Jurafsky & Martin, “Speech and Language Processing,” 3rd ed. https://web.stanford.edu/~jurafsky/slp3/
[6] Christiano et al., “Deep Reinforcement Learning from Human Preferences,” 2017. https://arxiv.org/abs/1706.03741
[7] Ouyang et al., “Training Language Models to Follow Instructions with Human Feedback,” 2022. https://arxiv.org/abs/2203.02155
[8] Rafailov et al., “Direct Preference Optimization,” 2023. https://arxiv.org/abs/2305.18290
[9] Google Research, “Deep Learning Tuning Playbook.” https://github.com/google-research/tuning_playbook
[10] Sutton & Barto, “Reinforcement Learning: An Introduction,” 2nd ed. http://incompleteideas.net/book/the-book-2nd.html; David Silver, UCL lectures. https://www.youtube.com/playlist?list=PLqYmG7hTraZDM-OYHWgPebj2MfCFzFObQ
[11] Lilian Weng, “Lil’Log.” https://lilianweng.github.io/
[12] Elhage et al., “A Mathematical Framework for Transformer Circuits,” 2021. https://transformer-circuits.pub/
[13] Bricken et al., “Towards Monosemanticity,” 2023. https://transformer-circuits.pub/2023/monosemantic-features
[14] Neel Nanda, “Mechanistic Interpretability.” https://www.neelnanda.io/mechanistic-interpretability
[15] Murphy, “Probabilistic Machine Learning.” https://probml.github.io/pml-book/
[16] Shalev-Shwartz & Ben-David, “Understanding Machine Learning.” https://www.cs.huji.ac.il/~shais/UnderstandingMachineLearning/
[17] Eugene Yan, https://eugeneyan.com/; “Patterns for Building LLM-based Systems & Products,” 2023. https://eugeneyan.com/writing/llm-patterns/
[18] Anthropic, “Building Effective Agents,” 2024. https://www.anthropic.com/engineering/building-effective-agents
[19] Lilian Weng, “LLM-powered Autonomous Agents,” 2023. https://lilianweng.github.io/posts/2023-06-23-agent/
[20] fast.ai, “Practical Deep Learning for Coders.” https://www.fast.ai/
[21] Goodfellow, Bengio, Courville, “Deep Learning,” 2016. https://www.deeplearningbook.org/
[22] Stanford CS336, “Language Modeling from Scratch.” https://github.com/stanford-cs336/assignment1-basics